Grok/xAI and official OpenRouter API connections use a function-only tool bridge for caller custom and namespaced tools, with original identities restored on the client side and encrypted reasoning preserved. See API providers for this compatibility route and named connection presets.
This document describes the OpenAI Responses ingress. The same provider catalog also serves native Anthropic Messages clients, including Claude Code. Selecting an inbound protocol does not select an upstream provider.
The gateway translates the caller's instructions and tool catalog. It does not assign agent roles, choose workers, add task instructions, or provide agent-only tools. Its catalog rows carry an empty instruction template, and the desktop adapter blanks any template an inference host or old cache supplies. On the desktop, Codex 2 rows take the base prompt and model messages from the client's own native catalog row for the same upstream model, so Codex sends what it sends that native model (details). Claude, Copilot and other rows, and clients without a native catalog, get no model-specific base prompt. Explicit and resumed client instructions remain authoritative.
Client tools must be registered before the gateway can transport them. The gateway SDK falls back to direct exposure for declared-deferred dynamic tools when the selected model lacks tool search; it does not invent scheduling or app tools. See client tool reachability for discovery behavior, the desktop scheduling boundary, and the isolated executable checks.
Where the detail lives
The reference is split into three focused pages, each readable on its own. This page stays the current protocol overview and keeps every section anchor that other documents link to, so existing references keep resolving here.
| Page | Read it when you are changing |
|---|---|
| Responses protocol, streaming, and failures | the fields the gateway adds or preserves, item replay, stream terminals and heartbeats, failure classes and retry budgets, capacity admission, and the upgrade and internal-field hardening |
| Responses tools across routes | tool declarations and identities, Chat Completions tool-call shapes, repeated declarations, the tool search bridge, and hosted web search |
| Responses provider adapters and images | the Claude CLI adapter and its prompt cache, Codex subscription login refresh, Copilot live sessions, the ordinary Anthropic API route, and image inputs |
Repeated tool declarations
A long chat can declare the same tool more than once, in Responses lite additional_tools input items or again in the current top-level tools list, and the gateway merges those declarations into its own catalog instead of refusing the request. Identity, precedence, scope and the tool_declarations_merged counts are on the tools page.
Full detail: Repeated tool declarations.
Tool search on the Codex subscription route
Native Codex rows can send a tool_search tool that the shared-subscription backend refuses, so search is off by default and a codex-subscription provider can opt in to a function bridge with gateway tool-search on. What each route drops, bridges, publishes and logs is on the tools page.
Full detail: Tool search on the Codex subscription route.
Hosted web search
web_search is the one hosted tool the ChatGPT backend accepts through a shared subscription, so the Codex subscription route forwards it unchanged while every other route drops it. Probe results, catalog behavior and ownership are on the tools page.
Full detail: Hosted web search.
WebSocket refusals, internal fields, and control items
/v1/responses is served over plain HTTP only: an upgrade attempt with a valid bearer gets 426 with no body, and recent Codex internal transport fields and executor control items are removed once in the request handler before any provider sees them. /health publishes the behaviors this release guarantees.
Full detail: WebSocket refusals, internal fields, and control items.
Ordinary Anthropic API
Configured reasoning effort maps to output_config.effort, outbound requests carry up to four ephemeral prompt-cache breakpoints, and only end_turn, stop_sequence and valid caller tool_use turns complete.
Full detail: Ordinary Anthropic API.
Images
User messages and caller tool results accept input_image with an inline base64 image_url, verified by signature and encoding, limited to 8 MiB each inside a 64 MiB request. Remote URLs and uploaded file IDs are rejected. Per-route handling, the Claude item-block cache format and inputModalities are on the providers page.
Full detail: Images.